View metadata, citation and similar papers at core.ac.uk brought to you by CORE provided by University of Canberra Research Repository 2009 International Conference on Future Computer and Communication Comparison of Artificial Neural Networks with Logistic Regression in Prediction of Kidney Transplant Outcomes Fariba Shadabi Dharmendra Sharma Faculty of Information Sciences and Engineering Faculty of Information Sciences and Engineering University of Canberra, ACT, 2601, Australia University of Canberra, ACT, 2601, Australia [email protected] [email protected] good analytical alternative to logistic regression techniques [2], [4]. In this paper we discuss our experience of applying data Abstract— Predicting the outcome of a graft mining techniques for the purpose of predicting the outcome transplant with high level of accuracy is a challenging task. To answer the challenge, data of medical procedures or events. The case study described mining can play a significant role. The goal of this in this paper is from the kidney transplantation domain. In study is to compare the performances and features earlier works, we addressed some of the practical issues of an Artificially Intelligent (AI)-based data mining associated with the use of ANN in the prediction of kidney technique namely Artificial Neural Network with transplant outcomes [5], [6]. In this paper we compare the Logistic Regression as a standard statistical data performances and features of an ANN approach to logistic mining method to predict the outcome of kidney regression in predicting renal transplantation outcomes. transplants over a 2-year horizon. The methodology employed utilizes a dataset made II. RENAL TRANSPLANT CHALLENGES available to us from a kidney transplant database. The dataset embodies a number of important The first successful human organ transplantation was properties, which make it a good starting point for carried out on December 1954 by Dr. Joseph Murray, in the purpose of this research. Results reveal that in Brigham hospital, Boston. In this instance Richard Herrick most cases, the neural network technique from Northboro, Massachusetts was given a kidney from his outperforms logistic regression. This study healthy identical twin brother, Ronald. He survived another highlights that in some situations, different eight years before the original medical condition destroyed techniques can potentially be integrated to improve his new organ. the accuracy of predictions. With the advent of more sophisticated anti-rejection drugs and antibiotics, organ transplants from related or unrelated donors have become much more common. Over the last five Keywords-Logistice Regression, Neural Network decades, thousands of kidney and other vital organ grafts such as heart and liver transplants have been performed I. INTRODUCTION successfully by surgeons around the world. Data mining techniques can be employed to support Although anti-rejection drugs have helped to boost the clinical data analysis. The logistic regression model is a success rate of transplants, the biggest challenge that statistical data mining method, which has been used widely continues to face transplant patients and surgeons is the risk by researchers in medicine for many years [1]. Logistic of rejection of transplanted organs. Over the years, there has regression is popular mainly because it enables the been substantial research into methods to predict graft researcher to avoid the need for many precise assumptions outcomes and identify key parameters influencing the required by other regression methods. Recently AI-based success or failure of transplanted organs [7]. Successful data mining techniques such as Artificial Neural Networks kidney transplantation will not only extend the longevity (ANNs) have drawn the attention of computer scientists and and quality of life for the recipient but also reduce medical clinicians for intelligent information retrieval from clinical expenses and increase the access to donor kidneys by data sources [2], [3]. reducing the need for multiple kidney transplants in the one Both ANN and logistic regression techniques have the patient. ability to model non-linear relationships between dependent Until now most clinical prediction methods have largely and independent variables. However research shows that been focused upon the use of standard statistical models. with the growing power of ANN tools, ANN can often be a However, statistical techniques often do not provide 978-0-7695-3591-3/09 $25.00 © 2009 IEEE 543 DOI 10.1109/ICFCC.2009.139 adequate knowledge for solving highly complex clinical classifier is trained with each of the training sets. As a result prediction problems. ANN have the ability to provide good each classifier may produce a different prediction [16]. solutions in situations where large number of variables Bagging could offer a significant improvement in prediction contribute to an outcome but their individual influence is accuracy. It is especially useful for a classifier with poor weak and not well understood. Clinical data gathered from performance, or where a classifier has been presented with a patients who have undergone graft transplant surgery have small training sample set, or where small changes in the data this characteristic and are known to be complex [8], [9]. can result in large changes in the classifier predictions (low In this paper we compare a widely used ANN approach stability issue). The pseudo-code of a classifier ensemble known as Multi-layer Perceptron (MLP) networks with with bagging is shown in Table 1. logistic regression, with the challenge being to select the Table 1. The bagging approach right kidney from the available pool of organs for a particular patient, thereby maximizing the chances for the Given: Training set S (with n cases and their classes), learning algorithm L, number of successful transplantation. bootstrap samples T III. MATERIAL AND METHODS Process:1.Model Generation A. Dataset for i = 1 to T: First, confirm that you have the correct template for your Generate a new training set(a bootstrap paper size. This template has been tailored for output on the sample) with n cases, using random US-letter paper size. If you are using A4-sized paper, please drawing (with replacement) from S close this template and download the file for A4 paper Apply the learning algorithm L to the bootstrap format called “CPS_A4_format”. sample B. Artificial Neural Networks and logistic regression keep the resulting model M(i) for future use The inspiration of ANN came from the desire to simulate features of the brain, namely biological neural networks and 2.Prediction for a given test case learning systems, which show high power in pattern for i = 1 to T: recognition tasks and adaptability. ANN architectures are Predict class of case using M(i) generally divided into two categories namely feed-forward network (networks without any loops in their path) and Return class that appears most frequently feedback networks (networks with recursive loops). Different network architectures and training algorithms can affect networks capabilities and usually a lot of trial and IV. METHODOLOGY error experimentation is necessary to determine the optimal Before The following methodology was employed: network topologies and training parameters. Detailed 1. Pre-process the data set. This includes: extracting information about the foundations of ANNs can be found in the data from different tables, cleaning the data, [12], [13]. For the purpose of this study, multilayer, feed- transforming the nominal attributes into numeric forward networks were used to differentiate between attributes, and choosing the appropriate parameters successful and unsuccessful transplants. All neural network to be included in the dataset with the help of classifiers described in this paper were implemented using domain expert. the Delphi programming language. 2. Split the dataset for training and testing (with Binomial (or binary) logistic regression is an effective balanced distribution of success and failure cases). supervised learning method that can be used to estimate the 3. Perform classification using the MLP network and probability of a certain event occurring. MLP network coupled with bagging. C. Bagging 4. Perform classification using the logistic regression and logistic regression coupled with bagging. An ensemble consists of several individually trained 5. Assess and compare the predictive accuracy of the classifiers that are jointly used to solve a problem. The most classifiers. popular methods for generating training sets for classifiers are Bagging and Boosting. We conducted two main experiments: Bagging (bootstrapping aggregates) was originally proposed Experiment 1. Neural Networks by Breiman [15]. Bagging is a popular method for training A single MLP classifier: Here we constructed a component classifiers. This technique generates several single MLP classifier. This experiment is reported training sets using random drawing (with replacement) from in more details in our previous study [5]. the original training set. Consequently, in every new training set there are data points, which appear more than Neural network ensemble: In this experiment, we once while others, do not appear at all. Each individual employed the bagging strategy to generate different 544 training sets from the original training set, using rate also reached 76% with 89%-agreement among the random drawing with replacement technique. Here networks